Segmentation of the guidewire tip in 2D X-ray image of percutaneous coronary intervention is a prerequisite for clinic applications, but also challenging for the tiny structure, large deformations and few texture features of the tip. In this paper, tip segmentation based on encoder-decoder convolutional neural networks with different receptive fields is explored, and a median tip distance error of 0.491 pixels, F₁ score of 0.982 is obtained by a modified structure with atrous convolution.